AI Thunderbolt! US Tech Giants Hide $3 Trillion in 'Hidden Liabilities,' Five Times Annual Capital Expenditure

The capital spending spree by US tech giants on artificial intelligence (AI) is bringing a massive 'hidden bill' into market view.

According to The Wall Street Journal, nine tech firms—Alphabet (GOOGL-US), Meta (META-US), Microsoft (MSFT-US), Amazon (AMZN-US), Oracle (ORCL-US), NVIDIA (NVDA-US), Broadcom (AVGO-US), SpaceX (SPCX-US), and Advanced Micro Devices (AMD-US)—have collectively disclosed around $3 trillion in off-balance-sheet commitments in recent securities filings, with the vast majority related to AI infrastructure development.

This amount is roughly five times the $600 billion in combined capital expenditure these nine companies made over the past year, and about three times their total outstanding lease and long-term debt balances. More notably, these off-balance-sheet commitments increased by approximately 50% in just two months, rising from about $1.8 trillion.

Currently, these obligations consist of two main components: approximately $1.2 trillion in 'uncommenced leases' and around $1.9 trillion in 'purchase commitments.'

Under current accounting rules, these liabilities typically do not need to be recorded on the balance sheet until leases begin or products and services are delivered, resulting in significant off-balance-sheet spending commitments.

AI Infrastructure Expansion: $3 Trillion in Obligations Hidden Off the Balance Sheet

'Uncommenced leases' are one of the primary drivers behind the rapid expansion of these off-balance-sheet liabilities.

For example, Meta’s Hyperion data center under construction in Louisiana spans an area equivalent to about 1,700 soccer fields. Meta has signed a lease agreement starting in 2029, initially for four years with the option to extend up to 20 years. If the lease is terminated early, Meta has committed to covering any shortfall to bondholders.

However, Meta believes the likelihood of actually paying the guaranteed amount is low, so it has not recognized the liability on its balance sheet. The obligation can remain off-balance-sheet until lease payments officially begin.

Meta has disclosed approximately $347 billion in uncommenced lease commitments. According to The Wall Street Journal’s analysis, the total uncommenced leases across the nine companies amount to about $1.2 trillion—nearly four times higher than a year ago.

The other major source is 'purchase commitments.'

As data centers expand at scale, demand for hardware such as NVIDIA GPUs and memory chips required for AI model training and computation has surged. To secure future supply and lock in production capacity, tech companies often sign long-term procurement contracts with suppliers.

Under accounting standards, these commitments do not need to be recorded on the balance sheet until the products or services are delivered. The total purchase commitments across the nine companies currently stand at approximately $1.9 trillion.

In other words, the combined $3 trillion in uncommenced leases and purchase commitments represents a massive future expenditure behind AI infrastructure expansion.

RVG Emerges as Another Risk: Chipmakers Become the Final Guarantors

Beyond leases and purchase commitments, the AI financing market has recently seen another off-balance-sheet arrangement: 'Residual Value Guarantee' (RVG).

According to reports, the scale of such structures is currently around $70 billion. The mechanism typically involves a special-purpose vehicle (SPV) borrowing funds to purchase AI chips, with the financing backed by cash flows from usage contracts with AI companies.

If the AI company stops making payments, the chips can be re-leased or sold to repay the debt. If a shortfall remains after the sale, the guarantor covers the difference—and the guarantor is often the chip manufacturer.

Thus, companies like NVIDIA and Broadcom act as the ultimate backstop in some transactions, but these guarantee obligations are typically not directly reflected on their balance sheets.

Meta states in its filings that since the likelihood of the RVG guarantor making actual payments is low, it has not recognized any related liability.

Broadcom has applied this model in a transaction codenamed 'Big Sky.' Broadcom provides a guarantee for a $35 billion debt deal, with investors such as Apollo Global Management (APO-US) and Blackstone (BX-US) funding the purchase of custom AI chips, which are then leased to Anthropic. Due to the guarantee, the senior debt received investment-grade ratings, thereby lowering financing costs.

Bank of America strategists estimate that Broadcom’s AI XPV platform could accumulate around $370 billion in senior debt by mid-2029.

NVIDIA CEO Jensen Huang stated the company may provide up to 25% residual value support for relevant opportunities, emphasizing that each case will be evaluated individually, aiming to bring in more external capital while managing its own risk exposure.

Last week, NVIDIA also announced a $500 billion financing partnership with six US investment firms, including BlackRock and Goldman Sachs, partly to reduce the 'circular financing' problem where AI companies invest in each other and buy each other’s products.

Rating Agencies on Alert: Investors Struggle to Gauge True Leverage

As RVG and similar structures rapidly expand, rating agencies are beginning to scrutinize the potential credit risks posed by these off-balance-sheet arrangements.

Moody’s (MCO-US) warned that if such transactions increase significantly in a short period, they could become a major risk. The agency cautioned that Broadcom could face substantially higher obligations, potentially constraining financial flexibility and pressuring creditworthiness—even if its current debt leverage remains low.

S&P Global Ratings classifies Broadcom’s residual value support as a 'contingent debt-like obligation' and stated it will include such items in adjusted debt calculations.

For bond investors, the real issue lies in when and under what conditions these obligations could turn into actual losses.

Mariya Entina, portfolio manager at DoubleLine, believes these financial arrangements are, to some extent, exploiting the system to achieve higher ratings, describing the current market as entering a phase of higher financial engineering.

CreditSights analysts liken NVIDIA’s provision of residual value support to 'selling a put option.' In a booming market, such arrangements incur almost no cost; but if AI clients default and hardware market values decline, the importance of these guarantees rises sharply.

Brian Gelfand, co-head of global credit at TCW, also noted that these transactions differ from standard investment-grade credit underwriting, and due to their off-balance-sheet nature, they carry relatively higher tail risks.

However, the market is not entirely pessimistic. John Lloyd, global multi-sector and corporate credit head at Janus Henderson Investors, believes that triggering residual value support would require an extremely extreme market environment.

He pointed out that there is currently no sign of a cliff-like drop in token usage growth, and companies are not trying to hide contingent liabilities, but rather seeking financing through these structures.

Can AI Demand Support the Massive Bill? Market Outlook Splits

There remains a clear divide in the market over whether these off-balance-sheet commitments could evolve into systemic risks.

The optimistic view holds that demand for AI tools continues to grow rapidly, driving stock market gains and causing supply constraints for core hardware. From this perspective, current AI infrastructure investments are still supported by strong demand, and tech companies may be able to cover related expenses through future revenue and cash flow.

But another camp worries about a potential time lag between AI infrastructure investment and actual revenue.

First, some procurement commitments and leases are non-cancellable, meaning that even if AI revenue falls short of expectations, companies may still have to fulfill signed payment obligations.

Second, Chinese open-source models are offering near-par performance at a fraction of the cost of cutting-edge models. If businesses and consumers shift en masse to cheaper alternatives, the profit margins of hyperscale cloud providers could be squeezed—yet the already-signed infrastructure spending will not disappear.

Reportedly, related token prices have already fallen more than 50% over the past few weeks.

From a financial structure standpoint, another notable issue is the timing gap between capital expenditure, revenue, and free cash flow.

Both Alphabet and Amazon have recently experienced negative free cash flow, meaning capital expenditure has exceeded operating cash inflow. This suggests that hyperscale cloud providers may still need to rely on capital markets for funding over the coming period.

Meanwhile, much of the AI infrastructure is still under construction, and depreciation costs have not yet fully reflected in current earnings. Once these assets are completed and transferred to fixed assets, depreciation expenses could be concentrated in future financial reports.

Currently, these off-balance-sheet commitments do not constitute an imminent debt crisis, but the $3 trillion scale of future spending is prompting the market to re-examine the balance sheets of tech giants once seen as 'fortress-like.'